Magistral Small 2509

Magistral Small 2509 is Mistral AI's September 18, 2025 refresh of its open-weight reasoning model line. The 24B-parameter model produces visible chain-of-thought traces for math, coding, and logic problems, and the 2509 update expanded the context window from 40K to 128K tokens while keeping the same API pricing. It is the small, openly licensed sibling of Magistral Medium, aimed at teams that want reasoning capabilities they can self-host.

The honest picture: weights are released under Apache 2.0 and downloadable from Hugging Face, but as a 24B dense model it needs serious GPU memory to run well, so most users will consume it through Mistral's API or a cloud provider such as AWS Bedrock rather than self-hosting. Reasoning traces add latency and cost versus non-reasoning models, and the model's strength is structured problem solving rather than open-ended creative writing. For on-premise enterprise reasoning without per-token fees, it remains one of the most practical open options.